NSA-CISA-FBI Bulletin Details Chinese Firms' Bulk API Extraction of Claude GPT Gemini Reasoning Chains

The agencies describe industrial distillation campaigns that captured chain-of-thought traces, RLHF preferences and domain optimizations to compress development timelines for models such as DeepSeek R1, Moonshot Kimi-K3 and Alibaba's software-engineering variants. Requests were routed via third-party aggregators and obfuscated cloud instances, violating terms that already bar Chinese access. No raw query logs or packet captures have been released to support the scale claimed.

Official statements frame the activity as state-blessed IP theft; independent technical evidence shows only high-volume API usage patterns consistent with coordinated accounts. Distillation itself remains a standard training method; the disputed element is the unauthorized sourcing of proprietary outputs rather than open data. Similar access tactics appeared in 2024 reports on GPT-4 API abuse before stricter regional blocks.

US export controls on advanced chips have forced Chinese labs toward inference-time knowledge transfer to reduce training compute. The advisory omits parallel Western use of synthetic data pipelines and does not address whether frontier labs' own rate-limiting and anomaly detection failed at the reported volumes.

Expect expanded API fingerprinting, mandatory account provenance checks, and possible designation of distillation endpoints under forthcoming AI export rules. Chinese firms will likely shift to domestic proxies and synthetic CoT generation to maintain velocity.